Re: [RFC][RFT] memcg fix cgroup_mutex deadlock when cpuset reclaimsmemory

From: Daisuke Nishimura
Date: Wed Dec 10 2008 - 09:08:48 EST


On Wed, 10 Dec 2008 18:36:07 +0530
Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx> wrote:

> * Daisuke Nishimura <d-nishimura@xxxxxxxxxxxxxxxxx> [2008-12-10 20:53:37]:
>
> > On Wed, 10 Dec 2008 17:18:36 +0900
> > KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> >
> > > On Wed, 10 Dec 2008 16:41:26 +0900
> > > Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx> wrote:
> > >
> > > > On Wed, 10 Dec 2008 15:19:48 +0900, Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx> wrote:
> > > > > On Wed, 10 Dec 2008 10:49:47 +0530, Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx> wrote:
> > > > > > Hi,
> > > > > >
> > > > > > Here is a proposed fix for the memory controller cgroup_mutex deadlock
> > > > > > reported. It is lightly tested and reviewed. I need help with review
> > > > > > and test. Is the reported deadlock reproducible after this patch? A
> > > > > > careful review of the cpuset impact will also be highly appreciated.
> > > > > >
> > > > > > From: Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx>
> > > > > >
> > > > > > cpuset_migrate_mm() holds cgroup_mutex throughout the duration of
> > > > > > do_migrate_pages(). The issue with that is that
> > > > > >
> > > > > > 1. It can lead to deadlock with memcg, as do_migrate_pages()
> > > > > > enters reclaim
> > > > > > 2. It can lead to long latencies, preventing users from creating/
> > > > > > destroying other cgroups anywhere else
> > > > > >
> > > > > > The patch holds callback_mutex through the duration of cpuset_migrate_mm() and
> > > > > > gives up cgroup_mutex while doing so.
> > > > > >
> > > > > I agree changing cpuset_migrate_mm not to hold cgroup_mutex to fix the dead lock
> > > > > is one choice, and it looks good to me at the first impression.
> > > > >
> > > > > But I'm not sure it's good to change cpuset(other subsystem) code because of memcg.
> > > > >
> > > > > Anyway, I'll test this patch and report the result tomorrow.
> > > > > (Sorry, I don't have enough time today.)
> > > > >
> > > > Unfortunately, this patch doesn't seem enough.
> > > >
> > > > This patch can fix dead lock caused by "circular lock of cgroup_mutex",
> > > > but cannot that of caused by "race between page reclaim and cpuset_attach(mpol_rebind_mm)".
> > > >
> > > > (The dead lock I fixed in memcg-avoid-dead-lock-caused-by-race-between-oom-and-cpuset_attach.patch
> > > > was caused by "race between memcg's oom and mpol_rebind_mm, and was independent of hierarchy.)
> > > >
> > > > I attach logs I got in testing this patch.
> > > >
> > > Hmm, ok then, what you mention to is this race.
> > > --
> > > cgroup_lock()
> > > -> cpuset_attach()
> > > -> down_write(&mm->mmap_sem);
> > >
> > > down_read()
> > > -> page fault
> > > -> reclaim in memcg
> > > -> cgroup_lock().
> > > --
> > > What this patch tries to fix is this recursive locks
> > > --
> > > cgroup_lock()
> > > -> cpuset_attach()
> > > -> cpuset_migrate_mm()
> > > -> charge to migration
> > > -> go to reclaim and meet cgroup_lock.
> > > --
> > >
> > >
> > > Right ?
> > >
> > Yes.
> > Thank you for explaining in detail.
> >
>
> Sorry, I don't understand the context, I am unable to figure out
>
> 1. How to reproduce the problem that Daisuke-San reported
Ah.. sorry.

1) mount memory cgroup and cpuset.
(I mount them on different mount points, but I think this can also happen
even when mounting on the same hierarchy.)
2) make directories
2-1) memory
- make a directory(/cgroup/memory/01)
- set memory.limit_in_bytes(no need to set memsw.limit_in_bytes).
- enable hierarchy(no need to make a child).
2-2) cpuset
- make 2(at least) directories(/cgroup/cpuset/01,02)
- set different "mems".
- set memory_migrate on.
3) attach shell to /cgroup/*/01
4) run some programs enough to cause swap out/in
5) trigger page migration by cpuset between 01 and 02 repeatedly.
I think Documentation/controllers/memcg_test.txt would help.

feel free to ask me if you need additional information.

> 2. Whether the patch is correct or causing more problems or needs more
> stuff to completely fix the race.
>
I should consider more to tell whether it's all right to release cgroup_mutex
under attach_task, but some more stuff is needed at least.


Thanks,
Daisuke Nishimura.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/